Reinforcement Learning and Control
Model-based Reinforcement Learning and Planning
Object-centric Self-supervised Reinforcement Learning
Self-exploration of Behavior
Causal Reasoning in RL
Equation Learner for Extrapolation and Control
Intrinsically Motivated Hierarchical Learner
Regularity as Intrinsic Reward for Free Play
Curious Exploration via Structured World Models Yields Zero-Shot Object Manipulation
Natural and Robust Walking from Generic Rewards
Goal-conditioned Offline Planning
Offline Diversity Under Imitation Constraints
Learning Diverse Skills for Local Navigation
Learning Agile Skills via Adversarial Imitation of Rough Partial Demonstrations
Combinatorial Optimization as a Layer / Blackbox Differentiation
Object-centric Self-supervised Reinforcement Learning
Symbolic Regression and Equation Learning
Representation Learning
Stepsize adaptation for stochastic optimization
Probabilistic Neural Networks
Learning with 3D rotations: A hitchhiker’s guide to SO(3)
Layered Optical Flow
Layered models allow scene segmentation and motion estimation to be formulated together and to inform one another. They separate the problem of enforcing spatial smoothness of motion within objects from the problem of estimating motion discontinuities at surface boundaries. Furthermore, layers define a depth ordering, allowing us to reason about occlusions.
In [
], we present an optical flow algorithm that segments the scene into layers, estimates the number of layers, and reason about their relative depth ordering using a novel discrete approximation of the continuous objective in terms of a sequence of depth-ordered MRFs and extended graph-cut optimization methods. We extend layer flow estimation over time, enforcing temporal coherence on the layer segmentation and show that this improves accuracy at motion boundaries.
In [], we extend the layer segmentation algorithm using a densely connected Conditional Random Field. To segment the video, the CRF can use evidence from any location in the image, not just from the immediate surroundings of a pixel. Additionally, the CRF drastically reduces runtime of the segmentation step, while preserving the high fidelity at motion boundaries.
PCA-Layers [] combines a layered approach with a fast, approximate optical flow algorithm. Within each layer, the optical flow is smooth and can be expressed using low spatial frequencies. Sharp discontinuities at surface boundaries, on the other hand, are captured by the layered formulation, and therefore do not need to be modeled in the spatial structure of the flow itself, allowing highly efficient layered flow computation.
We also use layered models in the treatment of motion blur []. In a dynamic scene, objects can move and occlude each other. Together with the nonzero shutter speed of the camera, this creates motion blur, which can be complex close to object boundaries; pixel values arise as a combination of foreground and background. Using a layered model allows us to separate overlapping layers from each other, making it possible to simultaneously segment the scene compute optical flow in the presence of motion blur, and deblur each layer independently.
Members
Publications